What this tool tells you
Thirty-odd models, four providers, three prices each. This narrows that field to the handful worth evaluating: set a budget, a context requirement and your input-to-output ratio, and the survivors come back ranked by what a representative request actually costs you.
It is a shortlisting tool, not a recommendation engine. Capability differences between the survivors will not appear in any price table, and no amount of filtering substitutes for running your own task against three candidates.
The ratio slider is the important control
Everything else is a filter. The ratio is what changes the ranking, and it is the number most people have never measured for their own workload.
- High ratio (10:1 and above) — summarisation, classification, extraction, retrieval-augmented answering. Input rate dominates and output pricing barely registers.
- Low ratio (2:1 and below) — content generation, code writing, translation. Output rate dominates completely, and the ranking from the high-ratio case frequently inverts.
Move the slider from 20 to 1 and watch the table reorder. That reordering is the entire reason a generic "cheapest models" list is not useful to you. Get your real ratio from production logs, or measure a representative request in the token counter.
What each filter does
Price caps
Hard ceilings on the published input and output rates. Useful for eliminating the flagship tier when you already know the task does not need it.
Minimum context
Filters on the verified window. Note that this excludes every model whose window we have not been able to confirm against the provider's own documentation — an omission rather than an invented figure, for reasons set out on the methodology page. If a model you expected is missing, that is why.
Exact token counts only
Restricts to models whose tokenizer can run in a browser, so pre-flight counts are exact rather than estimated. Worth switching on if you need to size prompts close to a context limit, or if your cost forecasting has to be defensible to someone else.
Hide legacy models
On by default. Superseded models often look attractive on price and are a poor place to start a new build, since they get deprecated on the provider's timeline rather than yours.
After the shortlist
- Take the top three or four into the cost calculator with your real monthly volume and cacheable share. A model that wins on a sample request can lose once caching is applied.
- Confirm your actual context fits in the context window calculator.
- Check your throughput against the quota in the rate limit calculator — a model you cannot get enough capacity on is not a candidate regardless of price.
- Then evaluate on your own task. Price differences are irrelevant among models that fail it.
One further note worth internalising: optimising the workload usually beats switching models. Before committing to a migration, read how to reduce LLM API costs — the first three items there routinely return more than any provider change.